Aiden：自主智能体你的浏览器控制终端持久记忆

Aiden — 自主智能体 AI agent operates 你的 computer 与 prompts: 浏览器 控制, 终端 execution, workflows, tools, recovery systems,、持久记忆. Built solo. AGPL-3.0.GitHub 468 ⭐ / 92 Fork，主要使用 TypeScript 开发。标签：agentskills,ai,ai-agents,ai-os,ai-skills。访问项目：https://github.com/taracodlabs/aiden

项目介绍

Aiden — 自主智能体 AI agent operates 你的 computer 与 prompts: 浏览器 控制, 终端 execution, workflows, tools, recovery systems,、持久记忆. Built solo. AGPL-3.0.

核心功能

- Durable autonomy. Every task gets a job-card; Aiden verifies work before claiming success, resumes after restart, avoids duplicate sends/actions, retries recoverable failures, and stops for permissions. - Evidence-required subagents. Helper agents must return real proof — a file, test result, id, or similar artifact — and Aiden re-checks that proof before trusting the result. - Trust dial. Choose how much Aiden can do: Observer, Assistant, or Partner. Destructive actions and spending still ask, even at the highest setting. - Glass dashboard + operator TUI. Watch model, token, budget, elapsed time, and tool steps live. Queue a follow-up, redirect mid-task, or interrupt cleanly while Aiden works. - One-shot mode + polish. `aiden -q "..."` for scripting, Discord setup inside the terminal, central cross-platform path handling, and green CI across Windows, macOS, and Linux.

使用指南

npm install -g aiden-runtime aiden ```

That's it. Pick a provider (Groq is free + fastest), paste your key, and start chatting.

Want autonomous triggers?

``` export AIDEN_DAEMON=1 # PowerShell: $env:AIDEN_DAEMON = "1" aiden trigger add file --path ~/Documents/inbox --label "watch-inbox" --include "*.txt" aiden # boots the REPL + dispatcher ```

Drop a file in `~/Documents/inbox/anything.txt` and Aiden acts on it. The agent turn is visible via `aiden runs list`.

<br>

https://github.com/user-attachments/assets/7a66bc19-8b17-4b01-be85-3aa5945a1b3b

<br>

What's new in v4.13

Aiden can now take a job, verify it, and keep going safely.

- Durable autonomy. Every task gets a job-card; Aiden verifies work before claiming success, resumes after restart, avoids duplicate sends/actions, retries recoverable failures, and stops for permissions. - Evidence-required subagents. Helper agents must return real proof — a file, test result, id, or similar artifact — and Aiden re-checks that proof before trusting the result. - Trust dial. Choose how much Aiden can do: Observer, Assistant, or Partner. Destructive actions and spending still ask, even at the highest setting. - Glass dashboard + operator TUI. Watch model, token, budget, elapsed time, and tool steps live. Queue a follow-up, redirect mid-task, or interrupt cleanly while Aiden works. - One-shot mode + polish. `aiden -q "..."` for scripting, Discord setup inside the terminal, central cross-platform path handling, and green CI across Windows, macOS, and Linux.

What's new in v4.12

Aiden can now connect outward to MCP servers and operate the browser with more state.

- MCP client. Connect to external MCP servers over stdio and Streamable HTTP, with OAuth 2.0 discovery, dynamic client registration, PKCE, token refresh, reconnects, circuit breakers, and one-command curated installs. - Approval-gated MCP tools. External tools are model-callable, but gated by approval. Results are redacted and fenced as untrusted input. - Browser depth. Accessibility-tree snapshots, stable element handles, act-by-reference click/type/fill, stale-element re-resolution, visual screenshot-to-vision support, multi-tab/dialog handling, and local browser attach/launch. - Safety breadth. Cross-session search, posture-aware skill/tool narrowing, truncate-then-redact tool-output caps, per-session budget enforcement, honest execution-policy messaging, and exfiltration-aware browser guardrails. - Process manager + activity. Tracked spawns, reliable tree-kill, redacted process listing, shutdown reaping, plus `/home` and `/activity` commands.

What's new in v4.11

A reliability release for honest output, better control, and stronger setup.

- Honesty & verification. Per-turn outcome verifier catches failed tool results and contradictions instead of glossing over them. - Artifact registry. `/artifacts` tracks files Aiden creates, with provenance: source tool, originating request, and turn link. - Streaming & control. Streaming delta coalescer reduces terminal writes, while `/undo`, `/retry`, `/compress`, and `/usage` give direct operator control. - Task durability. Orphan-task boot sweep retires crashed-session tasks as interrupted; `/tasks all` shows tasks across sessions. - Setup polish. Wizard back-navigation and config detection fixes prevent working OAuth/env setups from being mistaken as empty. - Model updates. DeepSeek V4 Pro and V4 Flash added, model picker table improved, and weak-model UI gating reduces markup leaks.

What's new in v4.10

Aiden becomes inspectable, durable, and finally streams properly.

- Streaming default restored at the wizard layer. `display.streaming` now defaults to `true` for fresh installs — restoring what v4.1.4 attempted but couldn't operationally land because the wizard wrote the older default into your `config.yaml`. Existing users with explicit `streaming: false` keep their setting AND see a one-time per-session disclosure suggesting the flip. - Durable Task object. Every user prompt creates a persistent task row with status / goal / event back-references. `/tasks [active|completed|cancelled|<N>]` lists; `/adjust <task_id> cancel` or `/adjust <task_id> goal <new text>` controls. - `/trace recent` event inspection. Every tool call, permission decision, and task lifecycle event lands in a queryable `run_events` ledger. Filter by category / kind / name / tool_call_id; scope to the current run, this session, the last N hours, or all. - Bug D ghost-text cursor fix. Ghost suggestions now render on a footer line below the prompt; cursor lands right after your typed input. Third attempt that actually works — prior two (v4.9.2, v4.9.6) shipped inert because they tried to fight `@inquirer/core`'s cursor manager from inside the rendered string. - Permission picker scope clarity. `Session (this path)` / `Always (this command)` labels carry a dynamic qualifier showing what the scope actually covers. No more "I picked Session, why am I being asked again?" surprises when the same tool fires with a different primary arg. - Eval harness v1. 20 contract tests under `tests/v4/eval/` protect the v4.10 substrate from regression — PTY-driven where possible, mockProvider-backed for streaming guarantees.

What's new in v4.9.0

Aiden v4.9.0 ships three new feature families:

- Memory — three-namespace persistent memory (memory/user/project) with CLI + auto-review - Hooks — secure subprocess hook system (observe/decide/transform) with full audit trail - Strategic substrate — UUIDv7 IDs, W3C trace propagation, durable runs, idempotency, crash recovery

Plus theme system, MCP integration (Claude Desktop / Cursor / VS Code), and substantial UI polish.

See [CHANGELOG.md](./CHANGELOG.md) for details.

<br>

What's new in v4.8

Aiden's UI got an identity, and pasting actually works now.

v4.8.1 — Paste handling + visual polish

- Robust paste handling. Stateful parser across stdin chunk boundaries, 800ms watchdog for missing paste-end markers, degraded marker form normalisation, universal CRLF/CR → LF, and timing-based accumulation that catches line-by-line paste delivery. Works across Windows ConPTY, SSH without `-t`, tmux/screen passthrough, and VS Code's integrated terminal. Typed prefix is preserved when paste arrives mid-input. - Markdown tables render as proper grids. Proportional column-width distribution with header-floor minimums so short labels never wrap mid-word. Model nudged via prompt to prefer sectioned lists for very wide comparisons. - Version reporting reads runtime `package.json`. Boot card and `/version` no longer report stale numbers after `/update install`. - `/update install` cleanups. Dropped the Node 20+ deprecation warning. Install progress shown via sliding-shimmer indicator. Install-method detection broadened to Windows user-mode globals. - Loading indicator + spacing. Indicator aligned at col 2 matching other surfaces; conditional erase preserves clean 1-blank rhythm between turns. - Status bar timer glyph (`⌛`). Visible at all terminal widths (mid + compact tiers previously dropped it). - Single approval panel per `file_write` (no duplicate event-row + panel stack).

v4.8.0 — Events backbone + visual identity

- Watch the agent work. Seven `ui_*` events (`ui_task_update`, `ui_task_done`, `ui_command_result`, `ui_test_result`, `ui_approval_request`, `ui_toast`, `ui_artifact_created`) let the model emit structured progress signals instead of writing them as prose. The REPL paints them as inline rows separate from the reply text. System prompt nudges the model to fire events during multi-step work. - Aiden-native visual chrome. Orange `│` panel bars with asymmetric top-divider chrome (no closing corners). Token-sourced design system (`cli/v4/design/tokens.ts`) — palette, glyphs, spacing in one place. - Packed status bar. Cyan model name · amber token ratio · semantic-tiered context bar (`●●●●●●○○○○`) · purple turn counter · teal `⌛` per-turn timer · state dot. Progressive disclosure on narrower terminals. - Sliding-block loading indicator. 4-cell `█` segment slides L→R on a muted `─` track at 250ms/cell. - Onboarding refresh. 24-bit ANSI AIDEN banner, capability-bullet disclaimer, 10-cell progress bar in the loading sequence, rounded heavy-frame "Built solo" card. - Markdown polish. Code blocks get a top-divider with brand-orange language label. Bullet lists use `●`/`○`; task lists use `✔️`/`○`. - ChatGPT Plus + gpt-5 routing fix. Auxiliary cheap-LLM calls (`risk_assess`, `compression`, `session_summary`, `skill_describe`) route through Groq + `llama-3.1-8b-instant` by default, falling back to the parent provider only when Groq isn't configured. Fixes the v4.6.x bug where every auxiliary call returned 400 from the Codex backend.

<br>

What's new in v4.6

Aiden now spawns workers and learns from itself.

- Sub-agents. `spawn_sub_agent` runs a focused child with an isolated context + intersected toolset; `subagent_fanout` runs N children in parallel (ensemble or partition) with merge strategies (`all` / `vote` / `pick-best` / `combine`) and provider rotation across configured fallback slots. - Operator kill-switch. `/spawn-pause on|off|status` blocks new sub-agent spawning while in-flight children continue. Marker file at `~/.aiden/spawn.paused` so the state survives restart and is shared across REPL, daemon, and MCP runtimes. Optional reason field captured in the typed `SUBAGENT_SPAWN_PAUSED` error envelope. - Self-improvement loop foundation. TCE classifications + recoveries persist to two new SQLite tables (`failure_signatures`, `recovery_reports`); `/recovery list|show|clear` surfaces recurring failure patterns across sessions. - REPL parent-run lineage. Each REPL turn writes its own `runs` row; sub-agent children link back via `spawned_from_run_id`. `aiden runs list` hides children by default and shows a `(N children, M OK)` badge per parent; `--include-children` flips to flat view. - PlannerGuard opt-in. The keyword-based per-turn tool narrower is OFF by default in v4.6 (modern models pick well from the full catalog). Enable via `/planner-guard on` or `AIDEN_PLANNER_GUARD=1` for smaller local models. - v4.6.1 onboarding redesign. Fresh disclaimer screen, loading sequence, rich provider picker with live `/models` fetch, 3-step connection probe, success screen, and `/walkthrough` guided tour. Phase 2 also fixed an MCP-mode `subagent_fanout` regression that had silently broken in the v4.5 refactor.

<br>

What's new in v4.5

Aiden now wakes up by itself.

- Persistent daemon mode (opt-in). `AIDEN_DAEMON=1` boots a background service with a SQLite-backed trigger bus. File watchers, webhook endpoints, IMAP polling, and cron schedules all feed the same durable queue. - Autonomous trigger dispatch. When a trigger fires, a real `AidenAgent` turn runs end-to-end — same tools, same sandbox, same recovery pipeline as your interactive REPL. Surface the chain with `aiden runs show <id>`. - Execution guardrails + risk tiers. Filesystem allow/deny lists (defense-in-depth for the `file_` tools — not shell containment; a shell command can still reach the filesystem) + dry-run preflight, plus an optional Docker session backend that is* real (if still weakly hardened) process containment. Default on; flip live with `/sandbox on|off`, inspect the honest policy with `/sandbox status`. - State-aware browser depth. URL + DOM + iframe-tree capture before/after every browser tool call. Stale-ref auto-retry. Surfaces login / 2FA / captcha / consent blockers as structured cards. - Continuous error recovery (TCE). 16 failure categories classified per tool call. Smart retry with cooldown. Dead-letter for permanent failures. Recovery report enriches the REPL's capability card. - Live-flip slash commands. `/sandbox`, `/tce`, `/browser-depth`, `/daemon status`, `/suggestions`, `/update` — toggle every subsystem without restart. Choices persist to `config.yaml`. - Update notification + `/update` CLI. Boot prompt when a newer version is on npm. Skip-this-version persistence. Install-method detection (npm-global / npx / standalone).

Full v4.5 internals: [`docs/v4.5/`](docs/v4.5/) (overview, triggers, architecture, daemon on Linux/macOS/Windows, troubleshooting).

<br>

Core capabilities

<br>

Architecture

Aiden is a local-first agent loop: provider adapters feed a 90-turn ceiling, every tool call passes through verification + failure classification + recovery, and tool results stream back as structured envelopes the model can reason over. v4.5 added a SQLite daemon foundation alongside the REPL — file watchers and webhook endpoints write to a durable trigger bus, a single-worker dispatcher claims events, and each trigger fires a fresh agent turn keyed by a stable session id. v4.12 added outward MCP client support and deeper browser state; v4.13 added job-cards, restart-safe task continuation, evidence-required subagents, and a trust dial. Sandboxed execution (filesystem allow/deny + Docker session backend), state-aware browser observation, and continuous error recovery apply to daemon-fired turns identically to REPL turns.

Detailed diagrams + module map in [`docs/v4.5/architecture.md`](docs/v4.5/architecture.md).

<br>

https://github.com/user-attachments/assets/323c9aa7-959a-425a-a5b3-4bae2b1a14bc

Install + first run

Linux / WSL / macOS (one-line)

``` npm install -g aiden-runtime aiden # interactive setup wizard fires on first run ```

Windows (one-line)

``` npm install -g aiden-runtime aiden ```

If you hit native-build errors on Windows, install the [windows-build-tools](https://github.com/felixrieseberg/windows-build-tools) prerequisite first.

Try without installing

``` npx aiden-runtime ```

After pulling updates

``` npm install -g aiden-runtime@latest ```

Or, from inside a running session, the slash command: /update install

(Aiden also prompts you on boot when a newer version is on npm — see `/update auto off` to silence.)

Uninstall

``` npm uninstall -g aiden-runtime ```

To also wipe your local Aiden home (config, sessions, memory, skills):

```

项目信息

项目地址

GitHub 仓库：<https://github.com/taracodlabs/aiden>

> 完整 URL：[https://github.com/taracodlabs/aiden](https://github.com/taracodlabs/aiden)

如需查看完整源码、安装命令、使用示例、API 文档，请直接访问上方项目地址。所有信息均以 GitHub 仓库为准。